Skip to content

Unit 4: Capstone Project

← Back to Lab Portal

Design, train, test, and honestly document your own Teachable Machine model — pulling together ethics (Unit 1), hands-on training (Unit 2), and an understanding of what's actually happening underneath (Unit 3).

What to turn in

  1. A problem statement — a real, specific task your model solves (not a generic two-random-objects demo). State it in one or two sentences.
  2. A data plan — what classes you used and why, and a note confirming your training data follows Unit 2's rule: uploaded files, non-identifying subjects, no photos/audio of yourself or classmates.
  3. Your trained model, with at least one documented round of testing and iteration (Unit 2, Lesson 3) — what didn't work the first time, and what you changed.
  4. Your Advanced settings, explained — what Epochs, Batch Size, and Learning Rate you used, and why, using what you learned in Unit 3 (too few epochs underfits, too high a learning rate destabilizes training).
  5. An honest accuracy report — tested with new examples your model never trained on, not the training examples themselves.
  6. A limitations reflection — at least one real case your model would likely get wrong, and why, using Unit 1, Lesson 4's bias lesson: what's missing or underrepresented in your own training data?

Optional stretch goal

Export your model and connect it to something interactive beyond Teachable Machine's own Preview panel — for example, a webpage using the exported sample code (Unit 2, Lesson 5), or, if you want to push further, triggering a command in your personal MUD based on its prediction. This is genuinely optional: nothing in the required list above needs it, and it adds real complexity (API calls, extra code) that's a stretch on purpose.

Try it

Complete all six required items into one submission. If your first model works perfectly on the first try, that's a sign to pick a slightly harder problem — a capstone with nothing to iterate on or document as a limitation isn't demonstrating much.

Tip

The honesty of items 5 and 6 matters more than how impressive the model is. A simple model with a clear, accurate account of what it does and doesn't do well is a stronger capstone than an ambitious one reported as "it just works."